In recent years, the momentum behind malaria elimination has gathered speed with thirty-nine countries across the world now making progress toward malaria elimination. One of the key strategies is to shrink the global malaria map from the endemic margins inward 1. While some nations are committed to nationwide elimination, others are pursuing spatially progressive elimination within their borders. With support from international donors, the Ministry of Health in the Republic of Vanuatu has started to implement a malaria elimination programme in Tafea Province which is comprised of five islands, Fatuna, Aneityum, Erromango, Aniwa and Tanna (Figure 1). Interrupted malaria transmission has already been achieved on the island of Aneityum through the use of approaches such as mass drug administration and insecticide treated bed nets, and with the enthusiastic support of the local population 2. Tafea Province represents the most southern and eastern limit of malaria in the South West Pacific and thus a strategic starting point for elimination activities.

While the optimal strategy for elimination is being debated, possible distinctions between control and elimination efforts are the geographical scale and intensity of key interventions 3. During the control phase, interventions tend to be widely applied to the target areas, often with similar strategies between zones of varying endemicity. As the prevalence declines remaining transmission is increasingly restricted to specific geographical foci and more precision in the application of interventions is needed, with more intense targeting of resources to identify and eliminate the last remaining sites of transmission and/or resistance 4. An effective elimination campaign must be capable of identifying these foci. Mapping offers great potential and the greatest need for malaria maps is at the periphery of stable, endemic areas where there is often less empirical information regarding risks and intensity of infection 5.

The advancement of geographical information systems (GIS) and spatial statistics has greatly improved our understanding of malaria dynamics, including its dependence on ecological factors 56789101112. More recently, Bayesian geostatistics has been embraced for disease mapping with the advantage that both environmental covariates and spatial autocorrelation are able to be estimated simultaneously and full posterior distributions produced, which can be used to quantify uncertainties in parameters of interest (e.g. predicted prevalence of infection)13. Spatial prediction models have been used to produce malaria risk maps at national 11141516171819 sub-continental 202122 and global scales 2324.

The aim of this present study was to produce accurate, validated predicted prevalence maps for P. vivax and P. falciparum on Tanna Island, Vanuatu. Additionally, it is envisaged that the maps be used to spatially define an implementation strategy for malaria elimination in Tanna Island. The applicability of the methods and the implications of the results are discussed in the context of malaria elimination strategies, which are beginning to take shape.

The overall prevalence of P. falciparum and P. vivax infection was 1.0% (95% CI: 0.79-1.21%) and 2.2% (95% CI: 1.06-3.34%) respectively 25 with the spatial distribution broadly similar between the species (Figure 2).

Semivariograms (Figure 2) revealed spatial autocorrelation was a feature of raw prevalence of P. vivax and P. falciparum. Colinearity of 0.91 (measured using Pearson's correlation coefficient) was observed between elevation and proximity to coastline. NDVI and TIR were not statistically significant predictors and thus were not included in the final models. Exclusion of the four aforementioned villages did not affect the significance of environmental variables and thus a separate model was not deemed necessary.

Tables 1 and 2 present the results of the Bayesian geostatistical models for P. vivax and P. falciparum. For P. vivax, the model with distance to coastline as the fixed effect gave the best fit with better predictive ability and lower mean error and mean square errors values than the model with elevation as the fixed effect. For P. falciparum, the model with elevation as fixed effect gave the best fit. On the basis of the models and validation it was possible to predict the distributions of P. vivax and P. falciparum risk across the island (Figure 4 and Figure 5 respectively). The strong relationship between proximity to coastline and P. vivax gave a smooth risk map with endemicity classes following the contour of the coastline. The risk maps for P. vivax and P. falciparum include the upper and lower Bayesian credible limits and show the extent of uncertainty in predicted malaria risk in 2008.

The Bayesian prediction models clearly show that transmission is not homogeneous, with malaria risk displaying a predominately coastal distribution, concentrated within well-delimited foci or 'hot-spots'. While the presence of malaria foci on the island is supported by anecdotal evidence the presence of an inland 'hot-spot' in the Green Hill area is contentious. An entomological survey carried out on the island concurrently with the parasitological survey did not reveal any vector breeding sites in the Green Hill area and concluded that transmission is restricted to within 2 km of the coast. It appears that the Green Hill area, which is approximately 5 km from the coast, does not support the local vector and that malaria infections were contracted whilst inhabitants visited coastal areas. The Green Hill area focus requires further investigation, particularly with respect to patterns of mobility for work and recreation. In addition, a better understanding of the transmission patterns could be obtained from surveying Anopheles farauti breeding sites on the island during the dry season.

The risk maps have been used to define zones within which interventions are scaled and planned according to malaria transmission intensity. This will involve indoor residual spraying (IRS) on the coastal fringe but not the hinterland/middle bush area. Additionally the maps provide the base for the design of the surveillance strategy and will be fully implemented by targeted surveys which will in turn inform modifications in local implementation policies such as inclusion of a particular inland hotspot for IRS.

Whilst the maps of the upper and lower Bayesian credible intervals are important for scientific honesty (i.e. highlighting areas where the predictions are imprecise) they are also a useful tool for priority setting and thus aiding in the sound and rational deployment of interventions. The maps showing the lower Bayesian credible intervals are particularly useful in highlighting those locations in which there is high certainty of above-average malaria prevalence. Active surveillance within high-prevalence foci is a cornerstone for success in interrupting malaria transmission 30. The upper Bayesian credible intervals are useful in providing an indication of the maximum extent of malaria transmission or risk.

At the community level the risk maps have been highly successful in planning meetings. The importance of grassroots-level involvement and local ownership of the elimination goal is well recognized 2. The island has had a long history of malaria control and such maps are an effective way for local citizens to visualize and understand the strategy for the elimination programme.

Initial statistics analysis of environmental covariates revealed no association between malaria risk and NDVI or TIR. At the small geographical scale of our study (550 km²), NDVI and TIR are relatively homogeneous and thus not indicative of the presence or absence of malaria. NDVI has been positively associated with malaria prevalence throughout Africa 818223132 and the Middle East 14, but these studies covered wide geographical areas exhibiting substantial variations in climate. The relatively small size of the study area meant that the commonly used source of vegetation cover, Globcover 33, was not available at a small enough resolution to usefully define vector habitats of An. farauti.

The present study found P. vivax risk to be associated with proximity to the coastline, which has been identified by others as a desirable An. farauti habitat 3435. It is believed the relatively small number of P. falciparum cases detected on the island were not sufficient to yield conclusive results concerning associations with environmental variables. Thus despite both P. vivax and P. falciparum being supported by the same vector species slightly different relationships with environmental variables were observed.

As the need to define the limits of malaria distribution and to predict the distribution within these limits increases, the utility of ecological and climatic variables in aiding this endeavour have been questioned. While an issue that cannot be resolved within this brief discussion suffice to say that as vectors' biological niches continue to be determined by ecological and climatic factors (along with human related socio-economic factors) understanding and identifying these variables will continue to be important. However, at differing spatial scales and for different vector species the significance of ecological and climatic variables will vary. For example, increases in altitude are commonly associated with cooler temperatures and thus a less suitable environment for malaria transmission for much of Africa 363738, while in the forested hills of South-East Asia increased altitude has been found to be associated with increased forest cover and thus a more suitable transmission environment for the vector species in this ecological niche 3940. Looking at smaller spatial scales, variables such distance to local water bodies and a households' location with respect to the village periphery become important determinants of malaria risk 1012.

Accurate data on distribution of population across the island was not available at the time of this study but has been identified as a priority for the malaria elimination programme. However, the population is described to be is spread over the entire island with an inland plateau region more densely populated 25. Once more accurate population data is available a better understanding of disease burden is possible.

While these maps will serve as baseline maps for the elimination programme such intensive surveying (i.e. 76% of all children aged 2-12 years 25) will not be necessary to update the maps. As surveillance operations become more streamlined the routine data collected from peripheral aid posts can be used. Additionally, the data collected throughout the year can be adjusted for seasonality 24 to provide more accurate estimates of average annual prevalence.

The current study does come with some limitations, particularly the geo-referencing method, which was applied retrospectively using independently sourced coordinates (rather than during the survey using a global positioning system). Additionally, the cross-sectional survey design meant that prevalence data were obtained at a single time point and thus the data only represent a snap-shot of malaria risk, which is known to have important temporal dynamics. Despite these limitations, as data are updated with the input from surveillance operations, the spatial predictions presented here represent the baseline of a dynamic, and ideally shrinking, malaria risk map for Tanna Island.

The authors declare that they have no competing interests.

PacMISC, AMI and the VBDCP conceived the study and participated in its design and data collection. AMI assembled the bulk of the incidence data. AC and HR devised and implemented the analytical methods. AT provided processed satellite images for analysis. HR wrote the first draft of the manuscript. All authors participated in the interpretation of results and in the writing and editing of the manuscript.